Shield AI has made significant strides toward integrating human-piloted fighters and Autonomous Collaborative Platforms (ACPs), ensuring they can seamlessly operate together. The company recently conducted a series of tests where high-performance, autonomously controlled, jet-powered drones, executed formation flying and tactical maneuvers, paving the way for blended operations between crewed and uncrewed aircraft. These advancements are crucial for the future of air combat, where autonomous systems must collaborate effectively with human pilots.
In July 2024, Shield AI, in collaboration with partners Kratos and Parry Labs, took a major step towards this vision when they conducted dual-ship autonomy tests using two of Kratos’ MQM-178 Firejet drones. The flights were conducted near the Kratos factory in Oklahoma.
These Firejet flights offered a significantly more relevant glimpse into the future of human-machine teaming operations compared to traditional single-aircraft autonomy tests. “The Federal Aviation Administration [FAA] required us to have chase planes, in this case an L-39 and L-29, up in the air coordinating with the Firejets, essentially making this a four-ship collaborative human-machine team,” explained Justin Johnson, Shield AI program manager. “When you have two unmanned jet aircraft flying at 300-plus miles-per-hour with two manned jets in a very tight, restricted environment in terms of airspace, trust in your artificial intelligence [AI] wingman takes on a whole different meaning.”
Combat drones like the Kratos XQ-58A Valkyrie have regularly flown alongside crewed fighter aircraft as manned-unmanned teams. However, these missions typically involve the human pilot in the fighter taking responsibility for collision avoidance. To achieve truly collaborative partnerships between crewed combat aircraft and drones, as envisioned for Collaborative Combat Aircraft (CCA) programs, the drones must be capable of maneuvering as part of a team, flying in tactical formations, and rendezvousing at will – all without risking airborne collisions.
The tests represented the latest use of Shield AI’s Hivemind product on an unmanned aircraft and marked the culmination of a three-phase flight test series effort. The project was designed to mature the company’s ability to deploy Hivemind onto third-party platforms and conduct complex flight operations in rapid succession.
“The Firejets were modified with Parry Labs’ EC Micro edge computer hardware, enabling Shield AI autonomy to run on the edge – meaning the computation occurred on the aircraft itself, rather than relying on centralized computing at a ground control station or tactical operations center,” Justin Johnson explained.
The EC Micro is a small form factor edge computing solution, designed to handle complex tasks such as autonomy, AI, machine learning, and navigation cross-banding directly on the platform, thereby minimizing the reliance on centralized ground control stations.
Shield AI served as the lead systems integrator, overseeing the entire project, managing the integration of computational payloads from Parry Labs, and ensuring that Shield AI’s autonomy software could run effectively on the Firejets.
Crawl. Walk. Get ready to run
Shield AI and Kratos adopted an incremental, progressive approach to capability demonstration. Initial flights focused on proving the efficacy of the platform control interface and the AI pilot. The team progressed from nothing to flight-ready in 120 days, to their first flight in just 170 days. From there, they flew every 30-60 days to rapidly advance the jet hardware, autonomy software, and flight operations, ultimately achieving collaboration between multiple autonomous aircraft.
“Our first flight of the series was an initial demonstration of exercising autonomy with the Kratos platform and exercising the control interface, which Kratos provided for us to have command and control over the Firejet. For that first test we were running [the AI pilot] on the ground and sending commands to the jet. We started out with the basics to make sure the commands were being received – things like heading, speed, altitude. Then we used orbit commands to keep the Firejet inside a notional mission operating area,” said Matt Maroofi, Shield AI’s Chief Engineer for the project.
The Firejet has traditionally been used as a target drone by the military, so Shield AI put their own mission computer onboard the platforms, running their software stack that they integrated with a comms-mesh network radio. This setup allowed the agents to ‘talk’ to each other and to the ground, enabling collaboration and execution of the mission objectives.
“We were working in a very tight, restricted environment in terms of airspace that was available to us, plus we were mandated by the FAA to have a chase plane for each Firejet for see-and-avoid as a safety precaution, maintaining visual line-of-sight of the Firejet from the time it takes off to the time it lands,” explained Maroofi. This test flight complied with the FAA’s Certificate of Waiver or Authorization (COA) requirements, which added safety constraints to the autonomy and limited the available airspace.
The capstone flight
The drones were launched and recovered from an airfield under human control, but once in open airspace, they switched to autonomous mode, with onboard autonomy agents piloting them through turning, joining, and formation flying – entirely without human intervention.
The mission commenced with a lead Firejet being rail-launched from the ground, piloted by a Kratos Remote Control Operator (RCO), who then flew the Firejet to the test airspace, accompanied by a chase aircraft. The tight airspace for the test meant avoiding populated areas, as the Firejets were not permitted to fly over such areas while under autonomous control per the FAA. The Firejets flew under RCO control to and from the airspace.
Once in the designated airspace, the chase plane took over monitoring the Firejet as it climbed out to the test area. The RCO relinquished control, and the autonomy was engaged. The launch crew then focused on the second Firejet, which was also launched under RCO control and followed the first jet into the airspace, accompanied by its own chase aircraft.
“The Firejets had no knowledge of the mission prior to the autonomy getting booted up. There’s no pre-scripted plan aboard saying, ‘you’re going to do this today.’ None of that information is stored on the jet prior to booting,” explained Madison Blake, Shield AI’s integration lead for the project. “When it boots up, it’s sent commands for what it’s supposed to do. As Firejet One was in the area first, it booted up and got sent a command from the ground saying, ‘I want you to hold at this location.’ As Firejet Two entered the airspace, its autonomy booted up and was commanded to join in formation with Firejet One, wherever it is.”
“The commands issued to the AI agents were designed to be loose. We don’t cheat, we weren’t saying we need to be in this exact scenario – it was a dynamic situation. If it doesn’t make it to the rejoin for any reason, it will trigger a replan if needed. Both aircraft are truly autonomous in the purest, most sci-fi sense of the word, collaborating with each other, receiving state information about each other, and the safety pilots are hands-off, watching the situation for safety. They’re not contributing anything directly to the Firejet movements – it’s true autonomy,” said Blake.
Dynamic coordination
Traditional autopilot systems use waypoint navigation to fly from point to point, typically not making decisions along the way. In contrast, Shield AI’s Hivemind Pilot autonomy can make dynamic decisions. “What our system does is that when we tell it to ‘aggregate,’ the jets come off a dynamic position, and they have to figure out their route so that they can come into line abreast with each other and into formation flight,” explained Holtzner.
While attempting to achieve a favorable position in the air for timing purposes – primarily to save fuel – the second Firejet can intercept the first wherever it is in the airspace. There are only a few locations where this can be done most efficiently, saving fuel and allowing the mission to proceed to other phases.
The Firejets, with their higher maneuverability, had turn rate restrictions due to the chase aircraft, so they couldn’t just snap into formation. The target altitude was also limited by the chase plane’s ability to climb and maintain pace with the Firejets. Once in formation flight, the Firejets moved on to mission tasks such as a Combat Air Patrol (CAP), flying a 180-degree offset pattern with each aircraft flying counterclockwise, 180 degrees out of phase from each other. All this planning was done dynamically onboard the agents, with their collaboration evolving from a leader-follower relationship.
To deconflict the two aircraft, as is done in military aviation, tactical formations that are altitude-stacked were used, maintaining that stack throughout the mission. “If one Firejet went up to 12,000 feet, the second jet went to 10,000 feet,” explained Blake.
An illustration of the autonomous coordination was observed during dynamic adjustments in formation flight. “At one point in the mission, Firejet One determined that it was a bit in front of Firejet Two. So, it automatically flew a lag maneuver to let the other agent catch up,” Holtzner noted. “We’ve even seen it where Agent One predicts a little bit too wide of a lag and then Agent Two will actually do the same to let Agent One catch back up.” Utilizing geometry to maintain formation is a common technique for jet aircraft.
What’s next?
Shield AI is taking a proactive approach to developing its AI pilot, and it is using company resources to accelerate the technology without the kind of constraints and timelines that can exist in government programs. “We are doing this work, and we are moving fast because we know this technology is crucial for defense,” said Justin Johnson.
Nathan Michael, CTO of Shield AI, emphasized the flexibility and adaptability of the Hivemind architecture, which successfully managed simultaneous flights of both Firejets and the company’s V-BAT drones in different locations on the same days in July.
“Our focus is on creating adaptable, platform-agnostic technology that meets current mission needs across a wide range of platforms,” Michael explained. “We’ve developed a system that applies sophisticated AI and autonomy to various platforms and mission sets, providing the flexibility our customers require, all with the objective of proliferating resilient autonomy for the world.”
Contact the author: jamie.hunter@teamrecurrent.io